Charm++ parallel programming system

[Thomas Albers <talbers AT binghamton.edu>]

Hello!

> There actually never was native x86_64 support for MX because InfiniBand
> and x86_64 took over the market at about the same time MX came out.

> Since it exists on other 64-bit platforms the port should be pretty
> easy, just copying conv-mach-mx.h and conv-mach-mx.sh to
> arch/net-linux_86_64 from another platform and figuring out what the
> right #defines are.

As I said, the idea is to use Open-MX instead of TCP/IP networking in
the hope that the lower latency translates into greater speed when
running NAMD.

I have:
#define CMK_USE_MX                      1
#define CMK_NETPOLL                     1
#define CMK_BARRIER_USE_COMMON_CODE     0

This fails immediately, with Open-MX versions 1.4.0 and 1.4.901:

ta@porsche
 ~/NAMD_2.8_Source/charm-6.3.2/tests/charm++/megatest $
./charmrun +p2 ./pgm
Charmrun> error 93620 attaching to node:
Socket closed before recv.

Perhaps more interesting is the behavior when using OpenMPI (1.4.3). The
test suite finishes most of the time, but sometimes one sees this:
ta@porsche
~/NAMD_2.8_Source/charm-6.3.2/mpi-linux-x86_64-smp-mpicxx/tests/charm++
$ mpirun --mca btl mx,sm,self -H
porsche,porsche,porsche,porsche,ferrari,ferrari,ferrari,ferrari,yamaha,yamaha,yamaha,yamaha,michelin,michelin,michelin,michelin,michelin,michelin
./pgm
...
test 36: initiated [multi marshall (olawlor)]
[16] Stack Traceback:
  [16:0] CmiAbort+0x58  [0x58c26a]
  [16:1]   [0x58d8a2]
  [16:2] CmiHandleMessage+0x2b  [0x58d5c7]
  [16:3] CsdScheduleForever+0x5c  [0x5911d6]
  [16:4] CsdScheduler+0xd  [0x59124e]
  [16:5]   [0x58cc1c]
  [16:6]   [0x58d1c7]
  [16:7] +0x6ac4  [0x7fa3a8942ac4]
  [16:8] clone+0x6d  [0x7fa3a6f793ed]
------------- Processor 16 Exiting: Called CmiAbort ------------
Reason: Converse zero handler executed-- was a message corrupted?

Does the behavior of Open-MX differ in some subtle way from Myrinet-MX?
How could I be helpful in tracking down the bug, and is it worth it?

Regards,
Thomas